A framework for developing Neural Networks in hardware accelerators

Description

This framework was part of the Diploma thesis titled "Architectures and Implementations of the Neural Network LeNet-5 in FPGAs".

The main goal of this thesis was to create a LeNet-5 implementation in an FPGA development board, but also form a reusable framework/workflow which can be modified to model and develop other Neural Networks as well.

The authors of this thesis were (in alphabetical order):

• Georgios Evangelou (Dipl. Eng. Electrical and Computer Engineering Department, Univ. of Patras, Greece)
  • Email: gevangelou [at] hotmail.com
  • LinkedIn profile
• Nick Roussos (Dipl. Eng. Electrical and Computer Engineering Department, Univ. of Patras, Greece)
  • Email: nroussos97 [at] gmail.com
  • LinkedIn profile

Duration of the thesis: May 2019 - July 2020

Instructions

Most code is commented or self-descriptive.

In order to replicate our workflow, start developing in the order of the folders' indices.

Frameworks used:

• Caffe/Ristretto
• Python 3.8
• Keras
• C/C++
• Vivado Design Suite (Vivado HLS, Vivado, Xilinx SDK)

Contents

The folders contained within this repository are the following:

• 0.Resources: The MNIST dataset used to train, evaluate and use our implementations. It can also be downloaded from the official website.
• 1.LeNet_in_Python: Our LeNet-5 implementation in Python and a script to estimate its average execution time per image.
• 2.LeNet_weights_extraction: Two-stage process of extracting the network's weights from the Python implementation to be later used by our C/C++ implementation.
• 3.LeNet_in_C: Our LeNet-5 implementation in C/C++ and various configuration files.
• 4.LeNet_Debugger: Some scripts that assisted in developments and debugging.
• 5.Excel_Visualization:
  • An excel document that assisted in developments and debugging.
  • Two excel documents containing the accumulative evaluation results of our C/C++ implementation.
• 6.Testbench: A script that generates files to be utilized by the Vivado HLS testbench we have created (see 3.LeNet_in_C).
• 7.LeNet_in_Ristretto: The LeNet-5 model in Ristretto with the corresponding quantization.
• 8.Optimizations: Definitions and reports of the different architectures created utilizing a variety of optimizations.
• 9.Zynq:
  • The IP produced using Vivado and Vivado HLS.
  • The SDK project produced.
  • The PC<->Board testbench script.
• A1.Diploma Thesis: The diploma thesis PDFs in both English and Greek.

This repository is work in progress. Please contact the authors for any issue.